In this article, I will be covering a recent study published in Nature on robots which can adapt like animals using sophisticated machine learning algorithms and I will discuss what it could mean for us, the common man on the street. For the wider audience, Nature is regarded as one of the most prestigious scientific journals, published since 1869, usually communicating only the cutting-edge research.

Machine learning

As a brief introduction, you might want to watch this video by Gary Sims on machine learning.

Machine learning is becoming a part of our daily lives, particularly through the use of search engines and smartphone applications, where the applications are developing a sense of intuition to predict what kind of services or information we are trying to find.

Robots which can adapt like animals

The article published in Nature, shows how scientists formulated a novel Bayesian learning algorithm, which enabled different kind of robots to recover from bodily injuries. Robots achieve this by performing a series of diagnostic experiments to understand how they should change their behavior to function like a brand-new robot. The digest of this research is nicely summarized in the video from Nature down below, and explains how the robot is mapping the results of its experiments to improve on its sub-optimal behaviour.

At 2:40 minutes, the robot moves like a wounded animal and almost resembles Arnold Schwarzenegger’s T-800 android robot crawling at the end of Terminator 2, after enduring severe bodily injuries. Thanks to the Bayesian optimization algorithm used by the authors, it will become possible for robots of the future to develop a sense of duty to accomplish their missions despite injuries they might suffer during the process. Of course, one should hope that the machines of the future will have a strong sense of duty for human friendly missions, which is emerging as a hot topic of debate (with inputs from people like Stephen Hawking).

An equally important finding of this research is that, the injured robot is shown to identify ways to move even quicker than its uninjured walking speed. This means that the robots of the future could also improve on their design specifications to remove parts of their bodies which they judge to be redundant to improve their efficiency.

Bayesian Learning

So, what is Bayesian learning? Although it may sound technical and sophisticated, Bayesian learning is a model of how people learn from their environments to make optimal decisions and it is based on the foundations of the statistical work done by Thomas Bayes, who lived in England in 18th-century.

It is based on the idea that people make decisions by combining their past experiences (called a “prior”) with the information they collect by observation and update their understanding about the current state of their environment (called a “posterior”). This model is quite intuitive and actually not so difficult to understand. Imagine you decided to quit your current job to work in a new environment. During the first week of your new job, you would be making observations about your new workplace which might have different implicit or explicit rules compared with the previous workplace. But until you feel like you have collected enough observations about your new workplace, you would also partially rely on your previous experiences to make executive decisions and depending on the accuracy of these decisions; you would update your understanding about the current workplace.

For example, a previous influential study showed that a Bayesian observer model makes decisions overlapping with real humans around 76% of the time, which is not so bad, considering that we are notorious for making sub-optimal decisions!

Potential implications for emerging mobile technologies

In the landscape of mobile technology, there are a number of ongoing developments which rely on similar machine learning algorithms in order to provide us with a better user experience. For example, Google’s Now on Tap feature, which is built-in to Android Marshmallow, combines the prediction capabilities of a computer algorithm with data about our search preferences to understand what kind of information we want to get next. Initial engagement with the “Now on Tap” feature seems to suggest that it is not 100% accurate all the time, but this is exactly where the machine learning algorithms will come into play, to improve accuracy by updating the intuitions of the software.

The second line of development is at the hardware front. For example, Google’s self-driving car and Japanese Softbank’s Pepper robot. When they are fully released for commercial use, these automatons, if they have a suitable price tag, may take over some of our daily responsibilities such as driving us to work or helping with house management.

The Bayesian learning algorithms highlighted in the published research would be crucial for these robots to find ways to recover from the potential damages, improving their longevity and reducing the overall maintenance costs for the users. Such abilities would be a must for consumers who are thinking about adopting these novel technologies, but concerned about their longevity, maintenance and price tag. The manufacturers will need to address these issues so that their novel products will be embraced by the market. Although we are probably used to upgrading our smartphones in 12-24 month cycles, longevity and autonomous damage recovery would definitely be an important marketing appeal for these larger and more expensive pieces of hardware.

Potential implications for disaster situations and space science

The final domain in which this robotics can help is in places which are beyond our reach, or for missions too dangerous for humans. In these domains, it is important for robots to have the ability to run diagnostic experiments and find adaptive ways to recover from injuries they might suffer while executing their jobs. For example, the robot that was sent to investigate the state of the reactor in the Fukushima Nuclear Power Plant, following the tragic earthquake and tsunami of 2011, failed to complete its mission because of extremely high levels of radiation (see the footage here), could it has benefited from these new learning algorithms?

What about space exploration to distant and hostile environments? The robots working in these far away places would benefit massively from employing these optimization algorithms. As some of you might already know, Google is promoting a lunar challenge for 2017 to put a rover on the moon and one of the competitors is under development by Audi:

Robots with such adaptive capabilities would help us to help us to push the boundary of the amount of data that we can collect from these ventures. Maybe one day we will have Google “Lunar” Maps, ahead of our potential colonization of our the Moon in the distant future; so that we will be able to study its surface and try to make the most accurate plans for such a day. Maybe along the way, we might also get lucky to have our hands on a a special “Lunar Launch” edition Nexus smartphone with design elements from the rover itself! Previously, Samsung released a limited addition Galaxy S6 Edge with the Iron Man theme; and release of more special edition smartphones would surely translate some of this excitement to the consumers, which would also help with raising public awareness for these cutting edge scientific ventures.

Although machine learning seems like a niche field studied by just a few people, it is in fact a growing area of research and one that will trickle down to touch all of our lives – both at home and far away. As research continues and being translated into the consumer market, we will see changes in our daily lives and we will look back in ten or maybe 20 years from now and wonder at how far things have progressed.